AI Weekly Digest: Claude Haiku 5.5, Mistral Large 4, GPT-6.1 Sol, OpenAI Math, Watermarking and Gemini Changes

LLM

• Claude Haiku 5.5 (Anthropic) — $0.10/$0.50 per million input/output tokens for prompts up to 100K. For longer requests, both prices grow fivefold.
• Mistral Large 4 (Mistral) — one trillion parameters, 52B active. Currently an API preview; weights are promised by the end of October.
• Ultrafast for Sol — an accelerated GPT-6.1 Sol appeared in Codex/Work on Pro for $500. In the API on short context, it costs $12/$60 per million input/output tokens.
• OpenAI math (OpenAI) — hundreds of preprints from an unreleased model after attempts to solve around 4,000 problems. Some proofs have been formalized in Lean; others may still contain errors.
• OpenAI watermarks — text marking can be enabled voluntarily in the API; in the EU, ChatGPT and Codex will add it in the coming weeks. Short texts and math are difficult to detect.
• Free Gemini (Google DeepMind) — starting October 9, Google is removing Flash and Pro from free accounts. Only Flash-Lite will remain. AI Plus will keep Flash; AI Pro and Ultra will keep all three models.

Generative models

• Odyssey-3 — generates a world from a prompt and reacts to actions in real time. You can walk as a cat, paddle a kayak, and change events on the fly; Flash is available in the browser.
• Nano Banana 2.1 — an update based on Gemini 3.6 Flash. Outputting a 1K image costs $0.0336 excluding input and thinking, but in the Edit Drift test the image still drifts after a series of edits.